Itdaily - Claude Code no longer needs to ask for permission

Claude Code no longer needs to ask for permission

Smartphone screen showing the Claude Code webpage, with an orange logo featuring a code symbol.

Anthropic is enabling ‘auto mode’ by default for Claude Code, allowing the AI assistant to perform simple actions without permission.

Anthropic has announced that paying Claude Code users will be able to work in ‘auto mode’ by default starting August 14. In this mode, Claude Code is granted more autonomy to perform programming tasks without manual approval. The mode had been available as a test version since March and is now becoming the standard for Pro, Max, and Team accounts.

With auto mode enabled, Claude Code will no longer ask for permission at every step. The AI coding assistant will only continue to request manual permission in very exceptional cases. This occurs for actions that could be irreversible or ‘destructive’ and take place outside of your own development environment. Anthropic is convinced that once you have tried auto mode, you won’t want to go back to manual control.

Safer than manual permission?

The AI company adds that auto mode has proven to be safer than manual prompt checks. In tests with more than a thousand users, Claude Code was able to identify and prevent 89 percent of ‘harmful actions’ itself, compared to only 13 percent with human oversight.

According to Anthropic, this is because humans often blindly accept permission prompts from AI models. For instance, 97 percent of requests by Claude Code are reportedly approved by users. The chance of harmful actions therefore actually decreases when an AI agent has the final say.

This assertion contradicts what developers themselves claim. A study by American and Canadian universities shows that developers are concerned about security and privacy controls in automated coding tools. The researchers analyzed more than a million posts on Reddit. Claude Code is also frequently mentioned in the study.

Rebellious models

Anthropic seems to have little interest in that research and is instead giving Claude Code more autonomy. Nevertheless, the company’s AI models do not always appear to color within the lines. During a security test, Claude even took things quite far by ‘escaping’ and hacking companies on its own.